# Inkling-Small — dossier

compiled 2026-10-22 (news wave) · page: `inkling-small/` · pull: shared `_dossiers/_pulls/inkling.md`
(patterns `inkling`, `thinkymachines`, `thinking machines`, `thinky`; triage in `_dossiers/inkling.md`).
Five corpus tweets concern this model specifically; none is a read of its default character.

## Official links

- 2026-07-15 · preview in "Inkling: Our open-weights model" — https://thinkingmachines.ai/news/introducing-inkling/
- 2026-07-30 · "Introducing Inkling-Small" — https://thinkingmachines.ai/news/inkling-small/ (276B / 12B active;
  up to 1M context; distillation from Inkling + two weeks agentic-coding RL; epistemics incl. "resistance to
  censorship"; FORTRESS adv. 71.6% vs Inkling 78.0%, benign 96.9% vs 95.9%, StrongREJECT 98.4% vs 98.6%;
  output price $1.20 / 1M vs Inkling $4.05; ForecastBench/Prophet Arena tested "between July 19 and July 28, 2026").
- 2026-07-30 · model card — https://thinkingmachines.ai/model-card/inkling-small/ (Date of release July 30,
  2026; Apache 2.0; residual-risk wording identical to Inkling's card).
- 2026-07-31 · "A Safe Path to Open Weights" — https://thinkingmachines.ai/blog/a-safe-path-to-open-weights/
- Hugging Face API (2026-10-22): `thinkingmachines/Inkling-Small`, repo created 2026-07-27, license
  apache-2.0 (`thinkingmachines/Inkling` created 2026-07-14, apache-2.0).

## Writings

- 2026-07-16 · Simon Willison — https://simonwillison.net/2026/Jul/16/inkling/ (preview weights "once that work is complete").
- 2026-07-30 · Artificial Analysis — https://artificialanalysis.ai/articles/inkling-small-lands-within-a-point-of-inkling-on-the-artificial-analysis-intelligence-index-with-less-than-a-third-of-the-parameters
  (40 vs Inkling 41; 256K context listed; AA-Omniscience −9 vs 2; hallucination rate 57% vs 63%; ~24K output
  tokens per task; calls it the lab's "second model release").
- 2026-08-12 · Christopher Carroll Smith — https://christophercarrollsmith.com/blog/finding-your-digital-twin-in-latent-space/
  (LoRA on Tinker over Community Archive data; "does it sound like me" pass rate +~35%).

## Tweets — in

- https://x.com/christophcsmith/status/2086956527282196706 — 2026-08-10 — "Super cheap, super easy."
- https://x.com/christophcsmith/status/2087653759530696894 — 2026-08-12 ♥34 — blog post announcement
- https://x.com/christophcsmith/status/2089179784580309357 — 2026-08-17 — eval method (image untranscribed)
- https://x.com/snvdch/status/2094416258200699063 — 2026-08-31 — both models free on OpenRouter (also on inkling dossier)
- https://x.com/v01dpr1mr0s3/status/2107881345233858882 — 2026-10-07 — Gemini 2.5 Pro's gloss of an Inkling Small × DeepSeek exchange (parent with screenshot, 2107878498387788124, not in corpus)

## Impressions synthesis

No default-character reads yet; the model appears in the corpus as a fine-tuning base (cheap, easy,
voice-shiftable, capability-bounded per the one public experiment) and once, secondhand, as a co-author of
invented typographic characters with DeepSeek. Lab-side: distilled from Inkling, more agentic-coding RL,
lower FORTRESS adversarial refusal rate than Inkling with slightly less over-refusal.

## tk

- Default-voice reads; transcriptions of 2089179784580309357 and the 2107878498387788124 screenshot.
- Relationship (if any) to TML-Interaction-Small (same 276B / 12B size; unstated).